A short shelf on the machines now reading, drafting and deciding around your business. Written so a beginner can start tonight, dated so a professional can trust it tomorrow. Twenty entries. You can finish the lot in an evening.
Tap each reader. One of them never asked.
One inbox, five readers. The question was never whether machines read your mail. It is which readers are on your side, and which ones told you.
Most businesses never decided to adopt AI. It arrived inside the tools they already pay for: the office suite, the CRM, the document reader, each one switched on by an update, a plan change, or a permission screen that stood between somebody and their afternoon. So the useful starting question is not should we use AI. It is already here. The useful question is an older and calmer one: who is reading, what can they touch, and who do they answer to.
That question does not need a technologist. It needs the same judgement you already apply to people: you would not let a stranger sort your mail without asking who they work for. The twenty entries below exist so you can ask exactly that, in plain English, of every machine in the building.
A note on numbers: this page avoids them wherever they cannot be trusted. Adoption surveys are mostly marketing, and scam losses are only ever what victims admit. Where a claim cannot stand up, we would rather teach the idea than decorate it.
Each entry says what a thing is, why it matters to someone who runs a business, and what careful firms do about it, where anything needs doing at all. Every entry carries the date it was last checked, because this landscape moves and a claim without a date is a guess. The shelf is kept deliberately small: when something new earns a place, something stale loses one.
Last full check: September 2026 · Changed since last check: the small business exemption card. Tranche two of the privacy reforms arrived in draft on 31 August and leaves the exemption in place for now.
Short of time? Read three: machine reading, payment redirection and the verification call.
Seven words that make every AI conversation easier to have.
Software reading your email and files is not new and not optional. Your mail provider scans every message, workplace AI indexes what you write, your software parses your attachments, and scammers read most attentively of all.
The question is not whether machines read your mail. It is which readers are on your side, and which ones told you what they do.
They count their readers. The ring at the top of this page is a fair start for most businesses.
A model is the engine that predicts words. An assistant is a model with manners: it answers when asked and then stops. An agent is a model with hands: it can browse, click, send and spend on its own.
Nearly every frightening AI headline is about an agent. Nearly every tool on your desk is an assistant. Knowing which one you are looking at is most of the risk assessment.
Training is teaching: your content is absorbed into the machine and shapes it permanently. Processing is reading: your content passes through, is worked on, and is gone.
Consumer AI tools may train on what you type. Proper business terms promise processing only, in writing. The same engine behaves differently depending on which door you entered through.
They ask every vendor one question: is our data used for training? A crisp no, in writing, or the data stays home.
Software with read only access can look but cannot touch. It cannot send, change, delete or move anything, no matter what it concludes.
It cannot is a stronger promise than it will not. When you grant any tool access, the difference between reading and writing is the difference between a window and a key.
An AI's working memory. What is inside the window, it knows. What is outside, it has never heard of, including what you told it yesterday.
This is why generic AI meets you as a stranger every morning, and why any tool claiming to know your business should be asked exactly where that knowledge is kept, and who else can reach it.
When an AI fills a gap with something fluent and false, delivered in the same confident voice as everything true it says.
The trouble is not that machines make errors; ledgers have always had errors. It is that these arrive polished, and polish is what we usually use to judge reliability.
In careful firms, anything that feeds a decision gets a second source or a human eye. Preferably both.
The metered line through which one piece of software uses another. Modern businesses do not buy AI in a box; they rent intelligence by the minute through an API.
The terms travel down that line. Business grade lines carry written promises about training and deletion. The same engine on consumer settings promises less. Where the line runs matters less than what the contract on it says.
Where small businesses actually get hurt, and the habits that stop it.
A second code from your phone whenever a password is used somewhere new. A password can be stolen from anywhere on earth. The second code, mostly, cannot. You will see it sold as MFA or two factor.
Most break ins do not break anything. They walk through one login that had no second lock on it.
They turn it on for email first, because email resets everything else. Then banking, then the main system.
A criminal watches invoices and settlements move through email, then sends our bank details have changed at the believable moment, from an address one character off the real one. The money leaves as usual. Only the destination changed.
This is the single most expensive thing that routinely happens to a small firm's inbox. Nobody can tell you how many businesses were caught last year, only how many admitted it.
One rule defeats it, and careful firms make it law: changed details get a phone call, to a number already held, never to one offered in the email.
Phishing is a net: thousands of generic fakes, hoping. Spear phishing is a spear: one email built from your world, your deals, your suppliers, your timing.
The spelling mistakes died when AI started writing the spears. The tell is no longer the grammar. It is the request: urgency, plus secrecy, plus changed details.
A backup is not what you save. It is what you can get back. Until a restore has been tested, you own hope, not a backup.
Extortion software prices its demand on your desperation. A tested restore is the difference between an inconvenience and a ransom negotiation.
They restore one file today, and put a quarterly repeat in the diary.
A password manager is a vault that lets one strong key hold hundreds of strong locks. The spreadsheet called passwords is the opposite: one weak key that opens everything.
Passwords reused across systems leak once and open everywhere. The vault removes the reason they were reused.
An account that keeps working after its person has left. Nobody watches it, so nobody notices what it does, or who else has learned its password.
It is the quietest door in the building, and the first one a professional intruder tries.
When someone leaves, closing their logins belongs on the same checklist as the farewell card.
Thirty seconds on the phone, to a number you already hold, before money moves anywhere new.
It is the cheapest security control ever invented, and it defeats the most sophisticated fraud AI can write. No technology purchase outperforms it.
What is moving right now, and how to look at it without flinching.
If your Microsoft 365 or Google plan has AI switched on, it can read whatever you and your staff can read: mail, files and chats, in order to answer questions and draft replies.
It inherits your permissions. If everyone can open the salaries folder, then so, effectively, can the assistant everyone can ask. Tidy the access first, switch on second.
Models small enough to run on a computer in your own office, so that content never leaves the room. Plainer and slower than the big engines, but private by geography rather than by promise.
For professions where confidentiality is the product, the trade is beginning to make sense, and the machines are improving quickly. One to watch rather than one to buy today, for most.
AI given the ability to act: browse, click, send, pay. Enormous promise, and the source of nearly every AI went rogue story, because hands turn a wrong guess into a done deed.
The right first question about any tool is never how smart it is. It is what can it touch, and what can it undo.
Before anything acts for the business, they have someone show them, in plain English, exactly what it is able to do without asking.
A few seconds of audio is enough to copy a voice convincingly, and the tools are now cheap and ordinary. The urgent call that sounds exactly like the boss may be neither urgent nor the boss.
Voices and faces are no longer proof of anything. Processes, not ears, have to carry the weight now.
They agree one office rule: unusual money requests are verified through a different door. A callback, a code word, a question a stranger could not answer.
Most modern software does its work on rented computers, often overseas. Your accounting package, your CRM and your email already work this way, and always did.
Is it overseas is the reflex question. Is it kept is the better one. Stored or passing through, training or processing, deletable on request: those three answers matter more than the map.
AI powered now appears on everything and means almost nothing. Three questions cut through: what does it read, what can it touch, and what happens to what it learns.
A vendor with crisp answers to all three respects you. A vendor with fog is also telling you something. The absence of an answer is an answer.
Australia decided against a dedicated AI law and instead regulates the technology through the laws that already exist: privacy, consumer, discrimination. That is not a reprieve. It means the obligations started before you did. Six situations cover most small businesses. Find yours.
Read this section the way it is meant: it explains what the law generally expects, in plain English, so you can ask your lawyer better questions. It is not legal advice, and how any of it applies to your business belongs to a professional. Law moves faster than any other entry on this page, so check the date and the source on every card before you lean on it.
The privacy principles follow the information, not the tool. If you collected it, you answer for where it goes, including into an AI product somebody else built. The regulator has said this plainly about commercial AI tools.
A complaint to the regulator, an investigation into how the information was handled, and the harder cost: explaining to a client where their file went. Whose tool it was will not be your defence.
Before any tool touches client information, they get two answers in writing: is it used for training, and can it be deleted. The vendor questions on the shelf above are also legal homework.
From 10 December 2026, if a computer program substantially helps make decisions that significantly affect someone, your privacy policy must say so. And discrimination law never stopped applying: a biased decision is unlawful whether a person or a model made it.
The transparency failure is the easy case. The hard one is a person turned down for something, asking why, and the honest answer being that nobody knows. In law the decision is yours, even when the reasoning was the machine's.
They list every place software helps decide things about people: applications, pricing, hiring. If the list surprises them, that is the point. A human should be able to explain every decision on it.
The consumer law judges the claim, not the author. If AI drafts something misleading about your prices, your results or what your product does, the words become yours the moment you publish them.
Misleading conduct complaints, and the fastest reputational bill in business: a public correction in your own name for a sentence you did not read carefully.
One rule for the office: nothing AI drafted is published until a human has read every claim in it the way a regulator would. Because one might.
Two clocks start. A suspected serious breach must be assessed promptly, within 30 days at the outside, and the regulator and the people affected notified where serious harm is likely. And since last year, individuals can sue directly for serious invasions of privacy.
A notification done late or badly often costs more trust than the breach itself. The new right to sue means the person whose data leaked no longer needs the regulator to act. They can come themselves.
They write the first hour down before it is needed: who confirms, who calls the lawyer, who tells the clients, in what order. A breach plan drafted on the day of the breach is not a plan.
Many businesses under $3 million turnover have historically sat outside the Privacy Act, but the walls are closing from two directions. Removing the exemption entirely is agreed in principle, but the next round of privacy reforms, released in draft on 31 August 2026, leaves the exemption in place for now. And since 1 July 2026, lawyers, accountants, conveyancers and real estate professionals have been anti money laundering reporting entities, and reporting entities were never exempt, whatever their turnover.
Building on the exemption is building on sand that is being carted away. The businesses caught out will be the ones learning the rules in the week they start applying.
They run the business as if covered now. Everything on this page is cheaper as a habit than as a retrofit, and no client grades your privacy against your turnover.
Your obligations to clients do not transfer to your vendors. If the CRM's AI leaks a file or a document reader mishandles identity papers, the client relationship, and the legal duty, still sit with you.
You wear the notification, the complaint and the apology first, then argue with the vendor about fault afterwards, under a contract somebody signed in a hurry.
Three questions in writing before any AI feature is switched on: what does it read, what can it touch, what happens to what it learns. A vendor who cannot answer has answered.
Office of the Australian Information Commissioner guidance on privacy and commercially available AI products, the notifiable data breaches scheme, the statutory tort for serious invasions of privacy and small business obligations. The Privacy and Other Legislation Amendment Act 2024, including the automated decision making transparency obligation commencing 10 December 2026, with analysis from Bird and Bird, White and Case and Ashurst. The Australian Government's National AI Plan, with analysis from Bird and Bird and the International Association of Privacy Professionals. Australian Competition and Consumer Commission guidance on false or misleading claims, and ACCC and Scamwatch reporting on payment redirection scams. Microsoft documentation on Microsoft 365 Copilot privacy and permissions and on Exchange Online Protection. Google Workspace documentation on Gmail phishing and malware protection. Anthropic and OpenAI published policies on business data and model training. Australian Cyber Security Centre guidance on multi factor authentication and backups. Reporting on AI voice cloning, including material published by the American Bar Association.
A daily signal to sharpen judgement, for professionals who act, not react.
See the Briefs